Re: [boost] Re: RFC: Multithreading design constraints

From: Beman Dawes (beman_at_[hidden])
Date: 2000-06-05 19:06:16

At 06:39 PM 6/5/00 -0000, Miki Jovanovic wrote:

>--- In boost_at_[hidden], jsiek_at_l... wrote:
>> It seems to me that having all these operations
>> (read/write/add/test&set/exchange) is a bit overkill. What's wrong
>> with just a plain old mutex (with lock/unlock).  Is there that
much 
>> of a performance difference to be gained?
>
>Possibly. atomic_traits class is a set of utilities built on top of 
>mutex. But this organization allows it to be optimised for
particular 
>platforms.
>
>> Oh, and along the lines of Steve's comment about signaling locks, 
>> etc. the notions of a semaphore and of a condition variable (both 
>> are in posix threads) are certainly important... probably
important
>> enough to be considered separately from mutexes.
>
>I strongly suggest we draw a line between basic promitives and 
>extended synchronisation. Basic primitives should in my view include

>only:
>

  - thread?

> - mutex
> - signal
> - atomic_traits
>
>mutex and signal are absolutely the most basic primitives, and 
>atomic_traits allow for optimisation of extended synchronisation 
>functionality.
>
>Once we are happy with the basic primitives, then we can work on 
>semaphores and conditinal variables. And the most beautiful part is 
>that we will already have a standard library, because all of these 
>extended classes can be implemented in terms of basic ones.
>
>If we keep it KISS, we will have a powerful foundation on top of
>which we can make a multithreading library as varied as we want.
Does 
>this make any sence?

It does to me.  I'm plowing through a couple of books on threading,
and that seems to be the basic approach both authors take, too.

Incidently, getting these primitives right includes an error handling
strategy.  The C++ Standard does not specify the operation of
exceptions across threads, so presumably exceptions can only be used
within a thread.  Would a signal be a useful mechanism to report
errors across threads?

--Beman